PaperlayerView exact revision on arXiv

Source-linked AI summary

Sequential Scenario-Specific Meta Learner for Online Recommendation

Zhengxiao Du, Xiaowei Wang, Hongxia Yang, Jingren Zhou, Jie Tang

arXiv:1906.00391v1cs.IRcs.LGstat.ML

TL;DR

Cold-start recommendation is difficult when scenarios have few interactions and existing algorithms rely on extensive observed data. s2Meta uses sequential meta-learning to adapt scenario-specific recommenders; it achieves the best results across three datasets against the reported baselines. The authors also note that recommender architecture performance can differ by dataset.

  • Problem

    Long-tailed, short-lived recommendation scenarios often lack enough feedback, while many existing algorithms rely on extensive observed data.

    Source: §1 · p. 1

  • Method

    s2Meta uses a sequential meta learner to control scenario-specific recommender initialization, updates, and early stopping.

    Source: §1 · p. 1, §3.2 · p. 3, §4.2 · p. 4

  • Results

    s2Meta achieves the best top-N recall results on three datasets versus shallow and deep cross-domain models, general recommendation, and a heuristic baseline.

    Source: §5.2 · p. 6

  • Takeaways & Limitations

    The authors identify automated recommender-architecture design as future work, noting that the same architecture can perform differently across datasets.

    Source: §6 · p. 8, §5.4 · p. 8

    The evaluation assumes access to meta-training tasks and uses meta-testing scenarios unseen during meta-training.

    Source: §3.2 · pp. 3–4

Abstract

from arXiv · show

Cold-start problems are long-standing challenges for practical recommendations. Most existing recommendation algorithms rely on extensive observed data and are brittle to recommendation scenarios with few interactions. This paper addresses such problems using few-shot learning and meta learning. Our approach is based on the insight that having a good generalization from a few examples relies on both a generic model initialization and an effective strategy for adapting this model to newly arising tasks. To accomplish this, we combine the scenario-specific learning with a model-agnostic sequential meta-learning and unify them into an integrated end-to-end framework, namely Scenario-specific Sequential Meta learner (or s^2 meta). By doing so, our meta-learner produces a generic initial model through aggregating contextual information from a variety of prediction tasks while effectively adapting to specific tasks by leveraging learning-to-learn knowledge. Extensive experiments on various real-world datasets demonstrate that our proposed model can achieve significant gains over the state-of-the-arts for cold-start problems in online recommendation. Deployment is at the Guess You Like session, the front page of the Mobile Taobao.

1 INTRODUCTION

Recommendation scenarios can be sparse and short-lived, making scenario-specific training difficult with limited feedback and manually tuned hyperparameters. The paper proposes s2Meta, which learns to initialize, update, and stop scenario-specific recommenders.

1 INTRODUCTION

  • Long-tailed, short-lived scenarios often lack enough feedback, while hyperparameter tuning requires substantial human effort and observations.Existing algorithms also rely on extensive data, and cross-domain cold-start methods still require many shared samples.Source: §1 · p. 1
  • s2Meta learns to instantiate recommenders with good generalization from limited training data through a sequential scenario-specific learning process.The paper combines few-shot learning and meta-learning to address cold-start scenario recommendation.Source: §1 · p. 1
  • The meta-learner initializes a broadly suitable recommender, flexibly fine-tunes it, and stops learning timely to avoid overfitting.Its policy is learned across existing scenarios and is intended to guide knowledge acquisition in new scenarios.Source: §1 · p. 2

2 RELATED WORK

The related work reviews context-aware recommendation, cross-domain recommendation, and meta-learning. It situates the paper among methods that use context, transfer information across domains, or learn across tasks.

2 RELATED WORK

  • The section organizes prior work around context-aware recommendation, cross-domain recommendation, and meta-learning.These are the three topics reviewed in the section.Source: §2 · p. 2

2.1 Context-Aware Recommendation

  • Context-aware recommenders predict in the user-item-context space, but the weights of different information types are domain-dependent and difficult to tune for cold-start scenarios.Context can be observable, such as time or location, or unobservable, such as user intention.Source: §2.1 · p. 2

2.2 Cross-Domain Recommendation

  • Cross-domain recommenders transfer information across heterogeneous users or items, but many methods require shared interactions, while feature alignment assumes domain-prior knowledge.The cited approaches extend matrix factorization or factorization machines, use nonlinear mappings, or design networks for cross-domain transfer.Source: §2.2 · p. 2

2.3 Meta Learning

  • Meta-learning uses information across tasks to guide rapid learning within each task, with metric-based and parameter-based approaches represented in prior work.Recommendation applications have focused mainly on algorithm selection; one user-cold-start approach directly outputs recommender parameters, limiting representation ability.Source: §2.3 · p. 2

3 PRELIMINARY

The preliminary section formalizes recommendation across scenarios and casts each scenario as a learning task with limited training interactions. It defines meta-training and meta-testing settings for learning and evaluating a scenario-specific recommender.

3.1 Problem Formulation

  • A scenario-based recommender uses user, item, and scenario sets, and its task is to rank items for users by follow-up behavior with limited scenario-specific training data.New promotions with sparse user-item instances are one stated example of this setting.Source: §3.1 · pp. 2–3

3.2 Scenario-Specific Meta Learning Settings

  • Meta-training tasks pair each scenario’s training and testing interactions, and the meta-learner is optimized on testing loss conditioned on training data.The paper builds these tasks by randomly splitting interactions in existing scenarios and specifies hinge loss for the objective.Source: §3.2 · p. 3
  • For a scenario, the meta-learner generates recommender parameters through initialization and flexible gradient updates, then the recommender is evaluated on unseen meta-testing scenarios.Meta-training and meta-testing scenario sets are disjoint.Source: §3.2 · pp. 3–4

4 THE PROPOSED FRAMEWORK

The framework combines a user–item recommender with a meta-learner that controls scenario-specific initialization, parameter updates, and stopping. The meta-learner is trained across scenarios to minimize testing loss after adaptation.

4.1 Recommender Network

  • The recommender maps user and item inputs through hidden interaction layers to produce a recommendation score.Its output layer computes the score from the final hidden representation.Source: §4.1 · p. 4

4.2 s2Meta

  • s2Meta automates initialization, update strategy, and early stopping, which are expert-managed challenges in few-shot scenario learning.The challenges include random initialization that can converge slowly or overfit, scenario-varying update hyperparameters, and overfitting from learning too long on small datasets.Source: §4.2 · p. 4

4.2.1 Parameter Initialization.

  • s2Meta initializes scenario parameters from shared global values, then uses history-conditioned LSTM gates and parameter-specific controllers for flexible updates.The input gate acts like a learning rate, while the forget gate can help the learner discard previous parameters and leave a poor local optimum.Source: §4.2.1 · pp. 4–5

4.2.3 Early-Stop Policy.

  • The learned stop controller predicts a stopping probability at each step because regularizers and small validation splits may not reliably prevent or detect overfitting.The controller is an LSTM that encodes historical information and uses a stochastic stop policy.Source: §4.2.3 · p. 5

4.2.4 Training of Meta Learner.

  • Meta-training minimizes expected testing loss after scenario-specific learning, using back-propagation for initialization and update parameters and policy gradients for stopping parameters.The method ignores higher-order derivatives in the meta-gradient to support flexible multistep gradient descent.Source: §4.2.4 · p. 5

5 EXPERIMENT

The experiments evaluate s2Meta on two public datasets and a large Taobao scenario dataset, comparing it with recommendation baselines and analyzing its learning process and recommender architecture. The reported comparisons cover performance across datasets and the contributions of the meta learner’s components.

5.1.1 Datasets.

  • Amazon and Movielens scenarios contain 100–1,000 items, while Cloud Theme contributes over 1.4 million clicks across 355 scenarios in a six-day promotion season.The datasets are Amazon Reviews, Movielens-20M, and Taobao Cloud Theme; the latter also uses users’ one-month purchase history before the promotion.Source: §5.1.1 · p. 6, Table 2 · p. 6

5.1.2 Baselines.

  • The baselines span popularity ranking, general-domain NeuMF, shallow cross-domain methods CDCF and CMF, and deep cross-domain methods EMCDR and CoNet.NeuMF is trained across scenarios; CDCF and CMF transfer cross-domain information, while EMCDR and CoNet use deep architectures.Source: §5.1.2 · p. 6

5.2 Performance Comparison

  • 9.41% (Amazon), 2.87% (Movielens), and 3.95% (Taobao) are s2Meta’s reported average improvements; the first two are against EMCDR and CoNet, respectively.s2Meta also achieves the best top-N recall across all three datasets against the listed cross-domain, general-domain, and heuristic baselines.Source: §5.2 · pp. 6–7, Table 3 · p. 7
  • Scenario-specific baselines generally outperform NeuMF on Amazon and Movielens, but often do not on Taobao; deep cross-domain methods outperform shallow ones by over 10%.The authors note that the usefulness of scenario information varies across datasets.Source: §5.2 · p. 7

5.3 Process Sensitivity Analysis

  • The process-sensitivity analysis examines how the meta learner controls scenario-specific learning, with the authors reporting that patterns on Amazon also hold on the other two datasets.The analysis includes component variations and a visualization of a typical learning process.Source: §5.3 · p. 7, §5.3.1 · p. 7, §5.3.2 · p. 7

5.3.1 Impact of Different Parts.

  • The complete meta learner significantly outperforms variants with random initialization, fixed learning rate, or fixed learning steps; random initialization hurts most.The authors suggest random initialization can require more steps and more easily become stuck in a local optimum; the effect of removing early stopping is relatively small.Source: §5.3.1 · p. 7, Table 4 · p. 7

5.3.2 Case Study: the learning process learned by the meta learner.

  • As training loss stops falling, stop probability rises; the authors report that test recall is stable then, while weights and biases change at different stages.For the last hidden layer, input gates remain low and forget gates high; weight matrices change mainly early, while bias vectors change mainly later.Source: §5.3.2 · pp. 7–8, Figure 2 · p. 8

5.4 Analysis of Recommender Architecture

  • The interaction architecture generally outperforms the mapping architecture, reaching a 6.71% relative improvement on Taobao, while mapping slightly outperforms interaction on Amazon.The authors note that mapping can be more efficient because mapped embeddings can be precomputed and matching optimized with advanced data structures.Source: §5.4 · p. 8, Figure 3 · p. 8

6 CONCLUSION

s2Meta applies meta learning to few-shot, scenario-specific recommendation and automatically controls learning to converge while avoiding overfitting. Experiments on real-world datasets demonstrate effectiveness against several baseline types; automating architecture design is proposed as future work.

6 CONCLUSION

  • s2Meta uses meta learning for few-shot scenario-specific recommendation, automatically controlling learning to converge and avoid overfitting.The conclusion frames this as addressing cold-start problems in some recommendation scenarios.Source: §6 · p. 8
  • Experiments on real-world datasets demonstrate effectiveness against shallow and deep, general and scenario-specific baselines.Source: §6 · p. 8
  • Future work would automate recommender architecture design because the same architecture can perform differently across datasets.Source: §6 · p. 8

A APPENDIX

This appendix section outlines the order of its deployment, implementation, and experimental-setting descriptions.

A APPENDIX

  • The appendix first describes deployment and implementation, then compared-method configurations, and finally settings for three datasets.Source: A APPENDIX · p. 10

A.1 Deployment

The model is deployed in the Guess You Like session on Mobile Taobao’s front page.

A.1 Deployment

  • Deployment is in Mobile Taobao’s front-page Guess You Like session.Source: §A.1 · p. 10

A.2 Implementation Notes

The implementation notes specify the recommender and controller architectures, embedding sizes, and training configuration.

A.2 Implementation Notes

  • The recommender uses three hidden layers and one output layer; embedding sizes are 128 for Amazon and Alibaba Theme and 64 for Movielens.Each hidden layer has half as many units as the last layer; update and stop LSTMs have hidden sizes 16 and 18, respectively.Source: §A.2 · p. 10
  • s2Meta uses standard SGD with learning rate 0.0001 and weight decay 1e-5, implemented with PyTorch 1.0 in Python 3.6.It runs on a single Linux server with 8 NVIDIA GeForce GTX 1080 GPUs.Source: §A.2 · p. 10

A.3.1 Code.

The appendix describes baseline implementations and parameter searches, then specifies dataset construction and evaluation settings for the cross-domain and Taobao experiments.

A.3.1 Code.

  • NeuMF uses the author-provided code, while CMF and CDCF use existing implementations; EMCDR and CoNet are implemented in PyTorch.The implementation adapts NeuMF's input and evaluation modules to the experimental settings.Source: §A.3.1 · p. 10

A.3.2 Parameter Configuration.

  • Hyperparameters for CDCF, CMF, EMCDR, and CoNet are grid-searched on the meta-training set, while NeuMF uses the author's suggested pre-training configuration.The search spaces vary by baseline; NeuMF uses an MF embedding size of 8 and MLP layers [64, 32, 16, 8].Source: §A.3.2 · p. 10

A.4 Experiment Setting

  • Experiments use Amazon leaf categories, MovieLens tags, and Taobao themes as scenarios, with source and meta-training/testing data defined separately.Amazon uses leaf categories from different first-order categories for source, meta-training, and meta-testing; MovieLens uses movies without tags as source and divides tags between meta-training and testing; Taobao uses a one-month pre-click purchase history as source and randomly splits themes. Evaluation uses scenarios with 100–1000 items and 64 shared-user interactions for training, with comparative methods using the same training data and evaluation set.Source: §A.4 · p. 10
Loading 1906.00391v1…